Socket::connect() spawns a ConnectBG job that it interrupts after a 5-second timeout. However, the job doesn't get interrupted on Linux (and SO_RCVTIMEO/SO_SNDTIMEO don't have an effect on connect timeouts in Linux). Thus, the system default is used as the effective connect timeout (as a function of net.ipv4.tcp_syn_retries).

This affects all intra-cluster connections made by mongod/mongos (and shell/tools) in failure modes where SYN packets destined to the remote server are silently dropped.

Reproduce with:

$ sudo iptables -A INPUT -i lo -p tcp --dport 27017 -j DROP
$ mongod --fork --logpath log
$ mongo